Yeah they’d have to it seems, but real time transcription isn’t free. Even late model devices with better inference hardware have limited battery and energy monitoring. I imagine it’d be hard to conceal that behavior especially for an app recording in the background.

WetBeardHairs@lemmy.ml mentioned that mobile devices use the same hardware coprocessing used for wake word behavior to target specific key phrases. I don’t know anything about that, but it’s one way they could work around the technical limitations.

Of course, that’s a relatively bespoke hardware solution that might also be difficult to fully conceal, and it would come with its own limitations. Like in that case, there’s a preset list of high value key words that you can tally, in order to send company servers a small “score card” rather than a heavy audio clip. But the data would be far less rich than what people usually think of with these flashy headlines (your private conversations, your bowel movements, your penchant for musical theater, whatever).